Max Kellermann

> but the performance wasn't really significantly better than bare tcmalloc That doesn't convince me much. Blackbox benchmarks are very valuable but not capable of quantifying the gains from small...

btw. right now this PR doesn't make anything faster because rarely ever can it combine several syscalls into one, and io_uring doesn't offload `sendmsg()` to worker threads .... unless one...

> BTW, FWIW, this is exactly the reactor model used by Seastar, which is in turn the event driven framework used by Crimson. What is? (There's nothing special with event-driven...

> While I'm not personally familiar with boost::asio, I'm curious about your perspective on Seastar's performance. I'd appreciate hearing your thoughts on potential areas for improvement. When I say Seastar...

> While Seastar indeed uses its own memory allocator to mitigate system allocator overhead I havn't seen that allocator yet, but whatever this allocator is about, it's not the kind...

> > This is called _every_ time somebody calls `.then()` on a `future` or every time somebody does anything with Seastar. > > I disagree, this piece of code is...

> I do. _Every_ means _every_. If you use the great quantifier, please honor the consequences. You are right, but you're still missing the point. > Even in debug builds...

> Hi Max, sorry for the late response. I was having this long post on my TODO for long time. Don't be sorry, I didn't expect or ask for a...

For those interested in io_uring performance, here's a Linux kernel patch series that is scheduled for Linux 6.15: https://lore.kernel.org/io-uring/[email protected]/ It reduces some of the performance problems that I found while...

Closing this PR because I'm not going to be able to finish this. There have been very long delays with many of my other PRs, and at the current rate,...